Executive Summary

Key Figures

Source: Ramp AI Index, TechCrunch (2026-11-06)

The four figures come from one tally but point along four different axes. The first two are growth rates, measured over different spans and over different things. The third is how narrow a base that spending comes from, and the fourth is the distance between firms. Reading the four cards on a single yardstick produces the wrong conclusion.

Corporate AI token spend grew 42% during July alone, then grew 1.5% across August and September together. Ramp described that stretch as two straight months of virtually no growth in token spend. The headline TechCrunch ran on November 6 came out of those two months.

Start with where the number is made. Ramp sells corporate cards and spend management software, and when an employee at a customer company swipes a card, the vendor that received the payment lands in the company's books. The Ramp AI Index pulls the AI-related payments out of that pile, drawn from transaction data at more than 85,000 American firms, and publishes them monthly. Unlike a survey that asks companies what they use, it counts only money that actually left.

The slowdown showed up on the spending side alone. In September the share of firms paying for AI reached 55.9%, up 0.62 percentage points on the month, with payments recorded at 43.8% of firms for Anthropic and 40.6% for OpenAI. More companies were buying and more tokens were moving, and only the total on the invoice held still.

Ara Kharazian, Ramp's head economist, doubts how useful that adoption rate still is. When AI was new and OpenAI had the market almost to itself, the question of what percentage had started paying mattered; now more than half of firms already pay for AI and many of them pay several vendors at once. Growth in new customers has visibly bent. Writing in the Ramp Economics Lab newsletter, he put it this way: "It feels like we're at the top of the S-curve, where the growth in the market will have to come from more intense adoption rather than more customers."

October turned the number back up. Token spend in the week ending October 23 ran 8.9% above the week ending September 25. Kharazian read that rebound as something that happened against conditions working the other way. Those conditions are the subject of section 2.

If usage grew far faster than spend, the price of a token must have come down in between. Divide the two growth rates for the fourth week of October and the drop works out to about 21% in four weeks.

The reason the price fell lies on the supply side. OpenAI cut API prices in half for GPT-6 Sol and GPT-6 Luna in September, and Anthropic cut its own prices at a cost of roughly $600 million in revenue growth. When both companies mark down at once, a company running the same work gets a thinner invoice without changing anything.

Buyers moved too. Kharazian described what a company learns as it gets further into AI: "You're more likely to use defaults that prevent usage of very expensive models for long-running tasks. You're more likely to impose some budgets on employees. You're more likely to offload some of the spend from standard and light models to even cheaper light models via routers." He summed up where that leaves the sellers: "You can use more AI, but use it for less. That's great for the customers…but it's not necessarily great for the model companies."

Where the money goes shifted as well. Anthropic's share of corporate token spend rose from 51.3% in the week ending September 25 to 60.5% in the week ending October 23, and OpenAI's fell from 44.2% to 35.1% over the same stretch. On usage the two moved the same way, Anthropic from 56.1% to 59.8% and OpenAI from 40.2% to 37.2%. The combined invoice was not thickening, and the two companies traded places inside it anyway.

Ramp wrote that figure into its October report. That one line sends you back through the growth rates above. A movement in the overall index may be the average of tens of thousands of situations, or it may be a few hundred situations copied across almost unchanged.

Monthly spend per employee shows the distance once you split it by group. In September the median firm paying for AI spent $15.70 per employee, the top 10% a median of $813.43, and the top 1% a median of $9,567.53. More than 600 times separates the middle from the top. Spending in the top 1% came down 3.1% from a July high of $9,873 to September's $9,567.53, and a good part of the overall slowdown is in that one line.

Ramp notes the point in its own commentary: because the top 1% represents a small group of firms, its spending estimate is more volatile than the median and top 10% estimates. A move among a few firms becomes the headline number directly.

That is also what Kharazian guards against. "Almost the entire spend on AI is from venture-backed, high-growth tech companies," he said. "That's fine as long as AI spend continues to rise, but if it starts to slow down because of something in the market or some other factor, they will all start to slow down simultaneously, and that could lead to other cascading effects." A group that shares its funding conditions converges even when each firm decides on its own.

What remains is whether the other 99% can fill the space. Kharazian draws a line there too: "If you look at the use cases today, it's not necessarily clear they will get up to that $10,000 per employee per month mark that the 1% is at," since "it's not necessarily true that a doctor's office or a car dealership will need to spend that much on AI." He is unconvinced by the personal assistant agents the industry names as its next engine, judging them not as economically productive as coding. "The era of hypergrowth may be over, and now that we know there may not be as much juice to squeeze out of the top 1%," he told TechCrunch.

Among public sources on corporate AI spending the Ramp AI Index is one of the few that counts real payments, which is why it is worth quoting. Three conditions are worth confirming first. Ramp has never hidden any of the three, and all three tend to fall away as the figures travel into news coverage.

One thing comes before the conditions themselves. The numbers the Ramp AI Index publishes do not belong to a single family. Adoption counts how many firms paid, spend counts dollars, and usage counts the volume of tokens. The three can move in different directions within the same month. ExplainX's review of the same index tells readers to ask, of any AI share statistic, which of the three it is and which population it covers. Most of the disagreement between dashboards starts there.

4.1A Two-Month Total and a One-Month Figure Are Not the Same Yardstick

The 1.5% covers August and September together, and the 42% covers July alone. Lay the two side by side, call it a 28-fold collapse, and a difference in span has turned into a measure of the fall. October's 8.9% compares four weeks with four weeks, so it sits next to July's 42% fairly, and it cannot sit next to the 1.5%. None of this says the slowdown was imaginary. It says the size of the slowdown is not settled by comparing these three numbers.

A further snag is that 1.5% appears twice in the same report with two meanings. One is the August and September spend growth above. The other says early October spend stood 1.5% above the previous peak in August, a fresh high. The first 1.5% means almost no growth and the second means a record. Quoted without its context, the same number will support two opposite articles.

4.2The Token Figures Do Not Count Every Ramp Customer

Anything involving usage or price rests on a different denominator from the index as a whole. Ramp writes that these metrics represent a sample of Ramp businesses that spend on AI tokens and use its Token Spend Management product, that the sample is not representative of Ramp's total AI spend volume, and that results are nonetheless directionally similar. The 37.1% and the 8.9% come out of that smaller sample.

The index as a whole carries a limit too. Ramp states that its results likely underestimate actual AI adoption and spend, because many businesses use free AI tools where no paid transaction occurs and because employees use personal accounts to get work done. The sample is American, and it leans toward firms that run corporate cards and spend management software. Explaining Korean corporate AI spending with these figures means closing that distance first.

4.3The Median Company Actually Spent More

A different reading of the same data has appeared. The newsletter Agent Autopsies argues that October's AI pullback was in the headlines rather than in the data, pointing out that among companies paying for AI the median firm's monthly spend per employee climbed from $11.38 in May to $14.74 in August and $15.70 in September, a 38% increase in four months, rising every month. The same piece notes that Ramp's index of the effective price per million tokens fell from $1.15 in March to $0.68 in September, a decline of 41%.

What that analysis asks for is a distinction. Moving to a cheaper vendor, a finance chief tightening budgets, shutting down a failed deployment, and a growth rate easing are four separate things, and coverage packs them into one and calls it a drop in spending. The author's point is that not one of the four means companies are paying less for AI.

Both readings come from Ramp's figures and both hold. Look at the top 1% and it is a slowdown; look at the median and it is a spread. Which one you get to say is settled by the denominator rather than by the data.

Of everything in this index, the router is what a team can act on right away. Companies that began sending the same work to a cheaper model made their invoices thinner, and one question has to sit in front of that switch. Which tasks does which model handle well enough? In an organization that cannot answer it, a router stops being a device that lowers cost and becomes a device that hides where quality fell.

The answer is not on a benchmark leaderboard; it is in your own data. How far a light model holds up on your company's documents, your customers' questions and your floor's logs differs by organization. Finding out means being able to sort tasks by kind, holding a label that decides right from wrong for each kind, and keeping a record that writes those decisions next to what they cost. With any one of the three missing, the only thing that changed after the model swap is the cost.

What a split of tasks by price band actually looks like is something this blog has examined once before. Vercel AI Gateway: Cheap Models Do the Work, Expensive Ones Take the Money followed a division of labor in which open-weight models handled 71.5% of tokens and took 23% of spend. The reverse case, where prices fall and the invoice grows anyway, is in Cheap Agents, Expensive Bills. Unit price and total are two separate axes, and how far apart they travel in a single month is written into this Ramp tally.

Anyone writing or receiving an AI cost report can settle three things in advance.

• Does usage appear next to the dollar figure? An invoice on its own cannot separate a price that fell from work that stopped.
• Do the two numbers being compared cover the same span? A two-month total placed beside a one-month figure turns a difference in span into a trend.
• Is it a mean or a median? When a few teams pay most of the bill, a mean turns their situation into the whole company's.

All three are questions about how to count rather than about which model to pick. This is also where Pebblous most often lands when it works to get data into a usable state. Decisions stall not because the numbers are missing but because nobody wrote down what those numbers used as a denominator.

Thanks for reading this far. The original tally is open at the Ramp AI Index, and the coverage is collected at TechCrunch. If you have moved work to a cheaper model inside your own company, we would like to hear what you used to decide where quality still held.

Primary Data

• 1.Kharazian, A. (2026). "Cracks in the AI Thesis, Part 3 — Ramp AI Index, October 2026." Ramp.

Press Coverage

• 2.Bellan, R. (2026, November 6). "Ramp data shows AI spending stalling just as Anthropic nears its IPO." TechCrunch.

Analysis & Critique

• 3.Kharazian, A. (2026, October 14). "Cracks in the AI Thesis, Part 3." Ramp Economics Lab (Substack).
• 4.Thakker, Y. (2026, October 14). "Ramp AI Index, October 2026: Anthropic Leads, OpenAI Closes In, and Spend Growth Stalls." ExplainX.
• 5.Kent, C. (2026, October 26). "In October "AI Pullback" Was in the Headlines. But Nowhere to Be Found in the Data." Agent Autopsies (Substack).